Choose a hosted model when your data can leave your boundary under a contract you trust and you want the provider to patch and scale. Choose self-hosted when code or findings must stay inside your environment, or you need control over versions. Either way, the risks differ rather than disappear.
What changes between hosted and self-hosted?
With a hosted model, a provider runs the model and you send prompts over the network. You inherit their data handling terms, their uptime and their update schedule. With a self-hosted model, you run the weights on your own infrastructure or a cloud account you control. You own patching, capacity, access control and the choice of model version. Neither one makes outputs correct.
Which risks come from the OWASP LLM list?
OWASP's 2025 list names ten risks. Three help frame this choice. Sensitive information disclosure matters most with hosted models, because prompts leave your boundary. Supply chain matters most with self-hosted ones, because you pick up weights, runtimes and dependencies from outside. Unbounded consumption applies to both: a hosted model can run up cost, a self-hosted one can exhaust your own capacity. Prompt injection, improper output handling and excessive agency apply to both equally, since they depend on how you wire the model into tools.
What is the data question?
Security work often touches source code, vulnerability details, secrets and incident notes. Classify what the model will see. Public advisory text is a very different input from unreleased source code. Write the rule: which classes may go to a hosted endpoint under which contract, and which must stay local. Check retention and training use in the provider's terms rather than assuming.
Constructed example: a team uses a model to triage scanner findings. Finding titles and CVE IDs are public, so a hosted model is fine. The same team then wants the model to read internal service code to judge reachability. That input is sensitive, so they route it to a model running in their own cloud account, and keep the hosted one for public data.
What about reliability and quality?
Hosted models usually give you the strongest available models and no capacity work, but the version can change and an outage becomes your outage. Self-hosted models give you a fixed version, offline use and predictable behavior, at the price of running GPUs, upgrades and evaluation. Test on your own tasks. Do not assume that a general benchmark predicts how a model does on your codebase.
Is a hybrid the answer?
Often, yes. Route by data class and task. Public data and low stakes summarization go to the hosted model. Sensitive code analysis goes to local. Design the routing as a policy, log which route each request took, and review exceptions.
Does Safeguard take a side?
No. Safeguard supports Full AI, Zero AI and bring your own model, including connecting your own models, with on-premises and local options, so the platform's work can run without an LLM at all or with the model you choose. That is a configuration choice for your team.
How to decide
- Classify the inputs the model will see and write the allowed routes per class.
- Read the hosted provider's data retention and training terms and file them with your vendor review.
- For self-hosting, list weights, runtime and dependency sources and scan them like any other software supply chain.
- Set quotas and alerts for cost or capacity abuse on both routes.
- Evaluate candidate models on a sample of your real tasks.
- Log the route per request and review it monthly.
Quick check
Is self-hosting always more private?
It keeps prompts inside your boundary, but you take on patching, access control and supply chain risk for the model and its runtime.
Can I use both?
Yes. Many teams route public or low risk data to a hosted model and sensitive code to a local one, under a written policy.
Sources
Last reviewed: 18 October 2026.
